Spaces:
Runtime error
Runtime error
File size: 9,918 Bytes
372785b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 |
import os
import torch
import torch.nn as nn
from transformers import (CLIPImageProcessor, CLIPVisionConfig,
CLIPVisionModel, SiglipImageProcessor,
SiglipVisionConfig, SiglipVisionModel)
from .videollama3_encoder import (Videollama3VisionEncoderConfig,
Videollama3VisionEncoderModel, Videollama3ImageProcessor)
class CLIPVisionEncoder(nn.Module):
def __init__(self, vision_encoder, args, delay_load=False):
super().__init__()
self.is_loaded = False
self.vision_encoder_name = vision_encoder
self.select_layer = args.mm_vision_select_layer
self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')
if not delay_load:
self.attn_implementation = getattr(args, 'mm_attn_implementation', 'flash_attention_2')
self.load_model()
else:
# uncertain whether flash-attention-2 is supported during inference phase.
self.attn_implementation = 'sdpa' # 'eager'
self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_encoder_name)
def load_model(self):
if self.is_loaded:
print('Vision tower is already loaded, `load model` call again, skipping.')
return
self.image_processor = CLIPImageProcessor.from_pretrained(self.vision_encoder_name)
self.vision_encoder = CLIPVisionModel.from_pretrained(self.vision_encoder_name,
attn_implementation=self.attn_implementation)
self.is_loaded = True
def feature_select(self, image_forward_outs):
image_features = image_forward_outs.hidden_states[self.select_layer]
if self.select_feature == 'patch':
image_features = image_features[:, 1:]
elif self.select_feature == 'cls_patch':
image_features = image_features
else:
raise ValueError(f'Unexpected select feature: {self.select_feature}')
return image_features
def forward(self, images, **kwargs):
images = torch.cat(images)
if type(images) is list:
image_features = []
for image in images:
image_forward_out = self.vision_encoder(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
image_feature = self.feature_select(image_forward_out).to(image.dtype)
image_features.append(image_feature)
else:
image_forward_outs = self.vision_encoder(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
image_features = self.feature_select(image_forward_outs).to(images.dtype)
return image_features
@property
def dummy_feature(self):
return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
@property
def dtype(self):
return self.vision_encoder.dtype
@property
def device(self):
return self.vision_encoder.device
@property
def config(self):
if self.is_loaded:
return self.vision_encoder.config
else:
return self.cfg_only
@property
def hidden_size(self):
return self.config.hidden_size
@property
def num_patches(self):
return (self.config.image_size // self.config.patch_size) ** 2
@property
def num_patches_per_side(self):
return self.config.image_size // self.config.patch_size
@property
def image_size(self):
return self.config.image_size
class SiglipVisionEncoder(nn.Module):
def __init__(self, vision_encoder, args, delay_load=False):
super().__init__()
self.is_loaded = False
self.vision_encoder_name = vision_encoder
self.select_layer = args.mm_vision_select_layer
self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')
if not delay_load:
self.attn_implementation = getattr(args, 'mm_attn_implementation', 'flash_attention_2')
self.load_model()
else:
# uncertain whether flash-attention-2 is supported during inference phase.
self.attn_implementation = 'sdpa' # 'eager'
self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_encoder_name)
def load_model(self):
if self.is_loaded:
print('Vision tower is already loaded, `load model` call again, skipping.')
return
self.image_processor = SiglipImageProcessor.from_pretrained(self.vision_encoder_name)
self.vision_encoder = SiglipVisionModel.from_pretrained(self.vision_encoder_name,
attn_implementation=self.attn_implementation)
self.is_loaded = True
def feature_select(self, image_forward_outs):
image_features = image_forward_outs.hidden_states[self.select_layer]
if self.select_feature == 'patch':
image_features = image_features
else:
raise ValueError(f'Unexpected select feature: {self.select_feature}')
return image_features
def forward(self, images, **kwargs):
images = torch.cat(images)
if type(images) is list:
image_features = []
for image in images:
image_forward_out = self.vision_encoder(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
image_feature = self.feature_select(image_forward_out).to(image.dtype)
image_features.append(image_feature)
else:
image_forward_outs = self.vision_encoder(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
image_features = self.feature_select(image_forward_outs).to(images.dtype)
return image_features
@property
def dummy_feature(self):
return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
@property
def dtype(self):
return self.vision_encoder.dtype
@property
def device(self):
return self.vision_encoder.device
@property
def config(self):
if self.is_loaded:
return self.vision_encoder.config
else:
return self.cfg_only
@property
def hidden_size(self):
return self.config.hidden_size
@property
def num_patches(self):
return (self.config.image_size // self.config.patch_size) ** 2
@property
def num_patches_per_side(self):
return self.config.image_size // self.config.patch_size
@property
def image_size(self):
return self.config.image_size
class Videollama3VisionEncoder(nn.Module):
def __init__(self, vision_encoder, args, delay_load=False):
super().__init__()
self.is_loaded = False
self.vision_encoder_name = vision_encoder
self.args = args
if not delay_load:
self.attn_implementation = getattr(args, 'mm_attn_implementation', 'flash_attention_2')
self.load_model(self.args)
else:
# uncertain whether flash-attention-2 is supported during inference phase.
self.attn_implementation = 'sdpa' # 'eager'
self.cfg_only = Videollama3VisionEncoderConfig.from_pretrained(self.vision_encoder_name)
def load_model(self, args):
if self.is_loaded:
print('Vision tower is already loaded, `load model` call again, skipping.')
return
# merge_size is set to 1 by default, because STAGE1, STAGE1.5, STAGE2 are trained with merge_size=1
# for stage 3, the merge_size is set to 2 by argments.
self.image_processor = Videollama3ImageProcessor.from_pretrained(self.vision_encoder_name)
# merge_size is fixed to 1 for STAGE1, STAGE1.5, STAGE2, STAGE3 in encoder and can be modified in connector.
self.cfg_only = Videollama3VisionEncoderConfig.from_pretrained(self.vision_encoder_name)
self.vision_encoder = Videollama3VisionEncoderModel.from_pretrained(
self.vision_encoder_name,
torch_dtype=args.torch_dtype,
attn_implementation=self.attn_implementation)
self.is_loaded = True
def forward(self, pixel_values, grid_sizes, merge_sizes, **kwargs):
image_features = self.vision_encoder(pixel_values, grid_sizes, merge_sizes)
return image_features
@property
def dummy_feature(self):
return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
@property
def dtype(self):
return self.vision_encoder.dtype
@property
def device(self):
return self.vision_encoder.device
@property
def config(self):
if self.is_loaded:
return self.vision_encoder.config
else:
return self.cfg_only
@property
def hidden_size(self):
return self.config.hidden_size
@property
def num_patches(self):
return -1
@property
def num_patches_per_side(self):
return -1
@property
def image_size(self):
return -1
def build_vision_encoder(vision_encoder_cfg, **kwargs):
vision_encoder = getattr(vision_encoder_cfg, 'mm_vision_encoder', getattr(vision_encoder_cfg, 'vision_encoder', None))
if 'clip' in vision_encoder:
vision_encoder = CLIPVisionEncoder(vision_encoder, args=vision_encoder_cfg, **kwargs)
elif 'navit' in vision_encoder.lower() or 'damovl' in vision_encoder:
vision_encoder = Videollama3VisionEncoder(vision_encoder, args=vision_encoder_cfg, **kwargs)
elif 'siglip' in vision_encoder:
vision_encoder = SiglipVisionEncoder(vision_encoder, args=vision_encoder_cfg, **kwargs)
else:
raise ValueError(f'Unknown vision encoder: {vision_encoder}')
return vision_encoder
|